(1) Field of the Invention
The present invention relates to a dialogue supporting system which supports person-to-person dialogue.
(2) Description of the Related Art
Automatic translation techniques are being developed for the purpose of supporting inter-lingual dialogue when ordinary people converse with locals during overseas travel. Among the machine translation techniques, the statistical machine translation method and the syntax transfer method are well known. In the syntax transfer method, a dictionary, grammar and the like are held, as a database, within an automatic translation apparatus, and an inputted sentence is analyzed using the dictionary and grammar. It is a method in which the sentence is converted to another language based on an understanding of its meaning.
In contrast, the statistical machine translation method holds a large number of pairings (referred to as parallel corpora) for a certain language and a language which is the conversion target. In the process of being used by a user, conversion trends, such as which word a certain word is often converted to, or which phrase in another language is often associated with a certain phrase and so on, or statistical characters become markedly apparent. In this manner, the statistical machine translation method is a method in which language conversion is performed based on the conversion trends and statistical qualities of words and phrases, without regard for grammar.
Among such automatic translation techniques, the syntax transfer method has the disadvantage in that maintenance of the dictionary, grammar, and the like, is troublesome. For example, when an unknown sentence comes up, dealing with the new sentence is not possible unless such sentence and words are analyzed and added to the dictionary by a person. In contrast, the statistical machine translation method has an advantage in that it is sufficient to merely increase the parallel corpora corresponding to such sentence. In other words, unlike the automatic translation techniques such as syntax transfer method and interlingua method, the statistical machine translation method virtually does not require manual analysis of language data and addition of additional information, and improvement of translation performance can be expected simply through parallel corpus addition. On the other hand, in the statistical machine translation method, optimal conversion is performed by following a probability formula and searching within the parallel corpora for an affiliation with which the probability value is highest. This process has an extremely high processing load and has the disadvantage of taking considerable time until the parallel translation is found. Recently, with the increase in CPU speed and main memory capacity, development of statistical machine translation methods based on a large number of parallel corpora is actively being carried out. Currently, the focus of research and development is placed on the optimization of language models and translation models, increasing decoding (translation processing) accuracy, and automatic evaluation of translation quality (see, for example, Patent Reference 1: Japanese Laid-Open Patent Application No. 2004-102946 and Non-Patent Reference 1: Y. Akiba, M. Federico, N. Kando, H. Nakaiwa, M. Paul, 3. Tsuiji: “Overview of the IWSLT04 Evaluation Campaign”, International Workshop on Spoken Language Translation (INTERSPEECH2004-ICSLPSatellite Workshop), 2004).